DOKK Library

A Stochastic Calculus Approach to the Oracle Separation of $\mathsf{BQP}$ and $\mathsf{PH}$

Authors Xinyu Wu,

License CC-BY-3.0

Plaintext
                          T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11
                                       www.theoryofcomputing.org

                                                         NOTE



    A Stochastic Calculus Approach to the
      Oracle Separation of BQP and PH
                                                      Xinyu Wu
                   Received February 7, 2020; Revised June 17, 2022; Published June 23, 2022




       Abstract. Recently, Ran Raz and Avishay Tal proved that in some relativized
       world, BQP is not contained in the polynomial-time hierarchy (STOCโ€™19). It has been
       suggested that some aspects of the proof may be simplified by stochastic calculus.
       In this note, we describe such a simplification.


1     Introduction
A recent landmark result by Ran Raz and Avishay Tal [8] shows that there exists an oracle ๐ด
such that BQP๐ด * PH๐ด . It has been suggested by several people, including Ryan Oโ€™Donnell,
James Lee, and Avishay Tal, that some aspects of the proof may be simplified by stochastic
calculus. We describe such a simplification.
    As Aaronson [1] points out, there is a classical correspondence between the relativized
complexity of PH and the size of bounded-depth, unbounded fan-in Boolean circuits (Furst,
Saxe, Sipser [6]). Using this correspondence, the oracle separation reduces to upper bounds on
the statistical difference between two distributions. Concretely, it suffices to show that there
exists a distribution ๐’Ÿ over {โˆ’1, 1}2๐‘ such that

ACM Classification: F.1.3, G.3
AMS Classification: 68Q15, 81P68
Key words and phrases: quantum complexity, stochastic calculus


ยฉ 2022 Xinyu Wu
c b Licensed under a Creative Commons Attribution License (CC-BY)                  DOI: 10.4086/toc.2022.v018a017
                                            X INYU W U

    1. For any ๐‘“ : {โˆ’1, 1}2๐‘ โ†’ {0, 1} computable by a quasipolynomial-size bounded-depth
       circuit,
                                                           polylog (๐‘)
                              |E[ ๐‘“ (๐’Ÿ)] โˆ’ E[ ๐‘“ (๐’ฐ2๐‘ )]| โ‰ค     โˆš       ,              (1)
                                                                 ๐‘
       where ๐’ฐ2๐‘ is the uniform distribution over {โˆ’1, 1}2๐‘ . The notation E[ ๐‘“ (๐’Ÿ)] means
       Exโˆผ๐’Ÿ [ ๐‘“ (x)]. In fact, a weaker upper bound of 1/(log ๐‘)๐œ”(1) would suffice for the oracle
       separation.
    2. There exists a quantum algorithm ๐‘„ that queries the input once and runs in ๐‘‚(log ๐‘)
       time, such that                                                
                                                                   1
                                    |E[๐‘„(๐’Ÿ)] โˆ’ E[๐‘„(๐’ฐ2๐‘ )]| โ‰ฅ ฮฉ           .                    (2)
                                                                 log ๐‘
For an explanation of why these two items suffice for the separation result, we refer to Raz
and Talโ€™s paper [8]. In their paper, Raz and Tal use a truncated Gaussian for ๐’Ÿ. Moreover,
they take ๐‘„ to be the Forrelation query algorithm (first introduced as โ€œFourier checkingโ€ by
Aaronson [1], and further analyzed in [2]). In this note, we will describe a construction of a
related but different distribution ๐’Ÿ based on Brownian motion, which simplifies certain details
of the analysis. The resulting analysis gives the same bounds as Raz and Tal, up to constant
factors.


2     Overview
2.1   Strategy to construct the distribution ๐’Ÿ
For convenience, we shall refer to the distribution called ๐’Ÿ in the Introduction as ๐’Ÿ 0. We shall
use ๐’Ÿ to denote an auxiliary distribution we shall use to construct ๐’Ÿ 0 (see step 2 below).
   We will first give an overview of the strategy to construct the distribution ๐’Ÿ 0. In this section
we will not formalize stochastic calculus concepts, deferring this to Section 3.
   The construction has two main steps:
  1. Use a stopped Brownian motion to define a distribution ๐’Ÿ on [โˆ’1/2, 1/2]2๐‘ .
  2. Round ๐’Ÿ to a distribution ๐’Ÿ 0 on {โˆ’1, 1}2๐‘ which has the same expectation on Boolean
      functions. (This step is identical with [8].)
   We will define the distribution ๐’Ÿ by describing how to sample from it.
   Let x๐‘ก be a standard ๐‘-dimensional Brownian motion, where โ€œstandardโ€ means its covariance
matrix is ๐ผ ๐‘ , the ๐‘ ร— ๐‘ identity matrix. Let y๐‘ก = ๐ป๐‘โˆšx๐‘ก , where ๐ป๐‘ is the ๐‘ ร— ๐‘ normalized
Walsh-Hadamard matrix (the W-H matrix divided by ๐‘). Finally, let B๐‘ก be the 2๐‘ ร— 1 column
matrix formed by x๐‘ก on top of y๐‘ก ; for typographical convenience we write this as B๐‘ก = (x๐‘ก , y๐‘ก ).

Proposition 2.1. B๐‘ก is a 2๐‘-dimensional Brownian motion with covariance matrix

                                           ๐ผ         ๐ป๐‘
                                                        
                                        ฮฆB ๐‘            .
                                           ๐ป๐‘        ๐ผ๐‘

                     T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                        2
         A S TOCHASTIC C ALCULUS A PPROACH TO THE O RACLE S EPARATION OF BQP AND PH

We defer the proof to Proposition 3.3.
   Let ๐œ€ > 0 and consider the random variable ๐œ defined by

                     ๐œ B min{๐œ€, first exit time of B๐‘ก from [โˆ’1/2, 1/2]2๐‘ } .                    (3)

    Let B๐œ denote the random variable obtained by stopping the Brownian motion B๐‘ก at the
random time ๐‘ก = ๐œ (for appropriately chosen ๐œ€, see the line before Equation (12)). Let ๐’Ÿ be the
distribution of the random variable B๐œ .
    Finally, use the method of Raz and Tal to round ๐’Ÿ to obtain a distribution ๐’Ÿ 0 on {โˆ’1, 1}2๐‘
such that the following holds for every Boolean function ๐‘“ : {โˆ’1, 1}2๐‘ โ†’ {0, 1} :

                                      E [ ๐‘“หœ(z)] = 0 E 0[ ๐‘“ (z0)] ,                             (4)
                                     zโˆผ๐’Ÿ           z โˆผ๐’Ÿ

where ๐‘“หœ denotes the (unique) multilinear polynomial ๐‘“หœ : โ„2๐‘ โ†ฆโ†’ โ„ that extends ๐‘“ . We shall
describe the method and prove Equation (4) in Proposition 4.2.
   Therefore, if ๐’Ÿ satisfies Equations (1) and (2), then so does ๐’Ÿ 0. Hence, it will suffice to
analyze ๐’Ÿ instead of ๐’Ÿ 0.

2.2   Sketch of the quantum algorithm
The quantum algorithm ๐‘„ used in Equation (2) is very simple: since ๐ป๐‘ can be implemented
by a quantum circuit of depth ๐‘‚(log ๐‘), computing hx, ๐ป๐‘ yi will only take a single query.
This value will have a large positive expectation when (x, y) is drawn from ๐’Ÿ but will have
expectation 0 when (x, y) is drawn from ๐’ฐ2๐‘ . This will be proven in Section 6.

2.3   Comparison with the proof of Raz and Tal
The proof here essentially follows the structure of RT, and uses many of the same technical ideas.
The main differences are in the proof of the lower bound against bounded-depth circuits, while
the quantum algorithm stays the same and has only a minor difference in its analysis.
    Our main contribution lies in simplifying some aspects of Sections 5 and 7 from [8]. Because
our ๐’Ÿ is defined to be bounded within [โˆ’1, 1]2๐‘ , there is no need to analyze a truncation
function applied to ๐’Ÿ, as in [8, Claims 5.2 and 5.3]. Theorem 4.4 reproves [8, Theorem 7.4], using
some ideas from [8, Claim 7.2]. [8, Claim 7.3] is replaced with Lemma 4.3, while the analysis of
the random walk in [8, Theorem 7.4] is replaced with an application of Dynkinโ€™s lemma. Also
using Dynkinโ€™s lemma, we directly prove Theorem 4.4 using bounds on the second derivatives
of the function ๐‘“ , instead of relying on Isserlisโ€™s theorem for moments of multivariate Gaussians.


3     Technical preliminaries
We briefly review some probability and stochastic calculus concepts. See for instance [7, Chapters
2.1 and 7] for more details. We shall use the common notation (ฮฉ, โ„ฑ , P) to denote a probability
space, where ฮฉ is the sample space, โ„ฑ is the ๐œŽ-algebra of measurable sets, and P is the probability

                     T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                       3
                                                 X INYU W U

measure. A random variable is an โ„ฑ -measurable function X : ฮฉ โ†’ โ„ ๐‘ . A random variable X
induces a distribution which is a probability measure on โ„ ๐‘ , defined by ๐œ‡X (๐ต) = P(Xโˆ’1 (๐ต)).

Definition 3.1. A stochastic process is a parametrized collection of random variables {X๐‘ก } ๐‘กโˆˆ๐‘‡
defined on a probability space (ฮฉ, โ„ฑ , P) and assuming values in โ„ ๐‘ .

    Typically, the parameter space is ๐‘‡ = [0, โˆž). For each ๐‘ก, we have a random variable
๐œ” โ†ฆโ†’ X๐‘ก (๐œ”). On the other hand, for a fixed ๐œ” we can consider the function ๐‘ก โ†ฆโ†’ X๐‘ก (๐œ”), a trajectory
of the process.

Definition 3.2. Let ๐พ be a positive semidefinite symmetric ๐‘ ร— ๐‘ real matrix. An ๐‘-dimensional
Brownian motion {B๐‘ก } ๐‘กโˆˆ[0,โˆž) with mean 0 and covariance ๐พ is a stochastic process characterized
by the following:
   (i) B0 = 0 almost surely.
  (ii) for ๐‘ข, ๐‘ก โ‰ฅ 0 the increment B๐‘ก+๐‘ข โˆ’ B๐‘ก is independent of the past, {B๐‘  } ๐‘ <๐‘ก .
 (iii) for ๐‘ข, ๐‘ก โ‰ฅ 0 the increment B๐‘ก+๐‘ข โˆ’ B๐‘ก is distributed as an ๐‘-dimensional Gaussian with
       mean 0 and covariance matrix ๐‘ข๐พ.
 (iv) almost all trajectories are continuous.
We say that B is a standard Brownian motion if ๐พ is the identity matrix.

   We refer to [7, Section 2.2] for a proof that such a stochastic process exists on some underlying
probability space (ฮฉ, โ„ฑ , P).
   We now make an observation.

Proposition 3.3 (Restatement of Proposition 2.1). Let x๐‘ก be a standard ๐‘-dimensional Brownian
motion, and y๐‘ก = ๐ป๐‘ x๐‘ก . Let B๐‘ก = (x๐‘ก , y๐‘ก ). Then B๐‘ก is a 2๐‘-dimensional Brownian motion with
covariance matrix
                                                ๐ผ ๐‘ ๐ป๐‘
                                                       
                                        ฮฆB                .
                                                ๐ป๐‘ ๐ผ ๐‘

Proof. We will check items (i)โ€“(iv) in the definition of Brownian motion. First, we note that
B๐‘ก = (x๐‘ก , y๐‘ก ) = (๐ผ ๐‘ , ๐ป๐‘ )๐‘‡ x๐‘ก , and so (i), (ii), and (iv) hold for B since they hold for x.
    Now we show (iii), that is, for fixed ๐‘ก, ๐‘ข, we want to show that the random variable
B๐‘ก+๐‘ข โˆ’ B๐‘ก = (x๐‘ก+๐‘ข , y๐‘ก+๐‘ข ) โˆ’ (x๐‘ก , y๐‘ก ) is distributed as a Gaussian with mean 0 and covariance
๐‘ขฮฆ. Using property (iii) of x, we see that x๐‘ก+๐‘ข โˆ’ x๐‘ก is โˆš          a Gaussian with mean 0 and covariance
๐‘ข๐ผ ๐‘ , so (x๐‘ก+๐‘ข , y๐‘ก+๐‘ข ) โˆ’ (x๐‘ก , y๐‘ก ) = (๐ผ ๐‘ , ๐ป๐‘ )๐‘‡ (x๐‘ก+๐‘ข โˆ’ x๐‘ก ) = ๐‘ข(๐ผ ๐‘ , ๐ป๐‘ )๐‘‡ ๐บ ๐‘ , where ๐บ ๐‘ is a standard
๐‘-dimensional Gaussian. Using the fact that for an ๐‘› ร— ๐‘‘ matrix ๐ด, the random variable ๐ด๐บ ๐‘‘
is an ๐‘›-dimensional Gaussian with mean 0 and covariance ๐ด๐ด๐‘‡ , we obtain that B๐‘ก+๐‘ข โˆ’ B๐‘ก is a
Gaussian with mean 0 and covariance ๐‘ข(๐ผ ๐‘ , ๐ป๐‘ )๐‘‡ (๐ผ ๐‘ , ๐ป๐‘ ) = ๐‘ขฮฆ.                                           

    We now define stopping times.

Definition 3.4. Let X = {X๐‘ก } ๐‘กโˆˆ[0,โˆž) be a stochastic process on (ฮฉ, โ„ฑ , P). A random variable
๐œ : ฮฉ โ†’ [0, โˆž) is a stopping time for X if for any ๐‘ก โˆˆ [0, โˆž), the event {๐œ โ‰ค ๐‘ก} is independent of
{X๐‘  } ๐‘ >๐‘ก . The stopped stochastic process X๐œ is a random variable defined via X๐œ (๐œ”) B X๐œ(๐œ”) (๐œ”).

                        T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                                4
         A S TOCHASTIC C ALCULUS A PPROACH TO THE O RACLE S EPARATION OF BQP AND PH

   In particular, stopping times may be applied to a Brownian motion to produce a stopped
Brownian motion. For example, any constant ๐œ0 B ๐‘ก0 is a stopping time; the stopped Brownian
motion has distribution B๐‘ก0 . The first time that B๐‘ก exits the cube [โˆ’1, 1]๐‘ , ๐œ1 B inf {๐‘ก โ‰ฅ 0 | B๐‘ก โˆ‰
[โˆ’1, 1]๐‘ } is also a stopping time. The minimum or maximum of any two stopping times is a
stopping time. In particular, ๐œ in Equation (3) is a stopping time.
   We will need to use the following fact about Brownian motion.

Proposition 3.5. Let B๐‘ก be an ๐‘-dimensional standard Brownian motion, and ๐œ be a bounded stopping
time. Then, E[kB๐œ k 2 ] = E[๐œ].

Proof. This follows from a few well-known facts about Brownian motion. First, kB๐‘ก k 2 โˆ’ ๐‘ก
is a martingale [9, Proposition II.1.2(ii)]. Given a bounded stopping time ๐œ, E[kB๐œ k 2 โˆ’ ๐œ] =
E[kB0 k 2 ] = 0 [9, Proposition II.1.4], and so E[kB๐œ k 2 ] = E[๐œ].                          
    The main stochastic calculus tool we will use is Dynkinโ€™s formula, which, for a function
๐‘“ : โ„ ๐‘ โ†’ โ„ ๐‘ , relates E[ ๐‘“ (B๐‘ก )] to the second partial derivatives of ๐‘“ .

Theorem 3.6 (Dynkinโ€™s formula, [7, Theorem 7.4.1]). Let B be an ๐‘-dimensional Brownian motion
with mean 0 and covariance matrix ๐พ, let ๐œ be a bounded stopping time, and let ๐‘“ : โ„ ๐‘ โ†’ โ„ be a twice
continuously differentiable function. The following holds:
                                                     ๏ฃฎโˆซ ๐œ ร•                             ๏ฃน
                             E[ ๐‘“ (B๐œ )] = ๐‘“ (0) + E ๏ฃฏ           ๐พ ๐‘–๐‘— (๐œ•๐‘–๐‘— ๐‘“ )(B๐‘  ) ๐‘‘๐‘  ๏ฃบ๏ฃบ
                                                     ๏ฃฏ                                  ๏ฃบ
                                                     ๏ฃฏ                                            (5)
                                                     ๏ฃฏ 0 ๐‘–,๐‘—โˆˆ[๐‘]                        ๏ฃบ
                                                     ๏ฃฐ                                  ๏ฃป

where ๐œ•๐‘–๐‘— = ๐œ•๐‘ฅ๐œ•๐œ•๐‘ฅ .
                  2

              ๐‘–       ๐‘—


    We will also require the following tail bound on Brownian motion.

Proposition 3.7 ([9, Proposition II.1.8]). Let B be a standard 1-dimensional Brownian motion. For
๐‘Ž, ๐‘ก > 0,                                           
                                        Pr sup |B๐‘  | โ‰ฅ ๐‘Ž๐‘ก โ‰ค eโˆ’๐‘Ž ๐‘ก/2 .
                                                                        2

                                            0โ‰ค๐‘ โ‰ค๐‘ก


4   Reduction to a Fourier bound
The main technical part of Raz and Talโ€™s proof [8] shows that, for a Boolean function ๐‘“ :
{โˆ’1, 1}2๐‘ โ†’ {โˆ’1, 1} computable by a bounded-depth, quasipolynomial-size circuit, and a
multivariate Gaussian distribution ๐’ต over โ„2๐‘ ,

                            | E[ ๐‘“ (trnc(๐’ต))] โˆ’ E[ ๐‘“ (๐’ฐ2๐‘ )]| โ‰ค ๐‘‚(๐›พ ยท polylog(๐‘)),                (6)

where ๐›พ is a bound on the (pairwise) covariances of the coordinates of ๐’ต, trnc truncates ๐’ต
so that the resulting random variable is within [โˆ’1, 1]๐‘ , and ๐’ฐ๐‘ is the uniform distribution
over {โˆ’1, 1} ๐‘ . This is based on the ๐‘˜ = 2 case of Talโ€™s fundamental result [10] that gives a

                          T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                    5
                                                      X INYU W U

polylog(๐‘š) upper bound on the the level-๐‘˜ Fourier coefficients of Boolean functions computable
by a Boolean circuit of bounded depth and size ๐‘š. We state Talโ€™s exact bound as Theorem 5.1
below.
    Another natural way of viewing a multivariate Gaussian distribution is as the result of an
๐‘-dimensional Brownian motion stopped at a fixed time. We can also build the truncation into
the stopping time. This allows us to use tools from stochastic calculus to analyze the distribution.
    We first recall the definition of restrictions of Boolean functions.

Definition 4.1 (restriction). Let ๐‘“ : {โˆ’1, 1} ๐‘ โ†’ โ„ and let ๐œŒ โˆˆ {โˆ’1, 1, โˆ—} ๐‘ . Let free(๐œŒ) be the set
of coordinates with โˆ—โ€™s. We define the restriction of ๐‘“ by ๐œŒ as ๐‘“๐œŒ : {โˆ’1, 1} ๐‘ โ†’ โ„, where ๐‘“๐œŒ (๐‘ฅ) is
๐‘“ evaluated at ๐œŒ with the coordinates of ๐‘ฅ replacing1 the โˆ—โ€™s in ๐œŒ.

  Henceforth, we also identify functions on a Boolean domain, ๐‘“ : {โˆ’1, 1} ๐‘ โ†’ โ„, with their
multilinear polynomial representations (or Fourier expansions)
                                                       ร•              ร–
                                            ๐‘“ (๐‘ฅ) =           ๐‘“ (๐‘†)
                                                              b             ๐‘ฅ๐‘– .                     (7)
                                                      ๐‘†โІ[๐‘]           ๐‘–โˆˆ๐‘†


    The following result has been extracted from the proof of Equation (2) in [8, Sec. 5].

Proposition 4.2. Let ๐’Ÿ be a distribution on [โˆ’1, 1]๐‘ . Let z0 โˆผ ๐’Ÿ 0 be sampled by first drawing z โˆผ ๐’Ÿ.
Then, independently for each ๐‘– โˆˆ [๐‘], we will set z0๐‘– = 1 with probability (1 + z๐‘– )/2 and z0๐‘– = โˆ’1 with
probability (1 โˆ’ z๐‘– )/2. For any function ๐‘“ : {โˆ’1, 1} ๐‘ โ†’ โ„, after identifying ๐‘“ with its multilinear
polynomial representation, we have

                                            E [ ๐‘“ (z)] = 0 E 0[ ๐‘“ (z0)].
                                           zโˆผ๐’Ÿ             z โˆผ๐’Ÿ

Proof. The proof follows from the Fourier expansion. First, fix ๐‘ง โˆˆ [โˆ’1, 1]๐‘ , and then draw
z0 โˆˆ {โˆ’1, 1} ๐‘ using the procedure above.

                                      ๏ฃฎร•           ร–         ๏ฃน  ร•          ร–
                        0                               0
                      ๐‘“        ๐‘ง]            ๐‘“             ๐‘ง         ๐‘“         E[z0๐‘– | ๐‘ง] = ๐‘“ (๐‘ง).
                                      ๏ฃฏ                      ๏ฃบ
                E   [   (z ) |    = E ๏ฃฏ      b (๐‘†)     z ๐‘–
                                                             ๏ฃบ=      b (๐‘†)
               z0 โˆผz                  ๏ฃฏ                      ๏ฃบ
                                      ๏ฃฏ๐‘†โІ[๐‘]       ๐‘–โˆˆ๐‘†       ๏ฃบ ๐‘†โІ[๐‘]       ๐‘–โˆˆ๐‘†
                                      ๏ฃฐ                      ๏ฃป
Taking the expectation of ๐‘ง over ๐’Ÿ, we infer Ezโˆผ๐’Ÿ [ ๐‘“ (z)] = Ez0โˆผ๐’Ÿ 0 [ ๐‘“ (z0)].                       

    We make some observations about Fourier coefficients. First, the Fourier coefficients of ๐‘“๐œŒ
        ๐‘“๐œŒ (๐‘†) = 0 for all ๐‘† * free(๐œŒ). We also have that
satisfy b

                                                 ๐‘“ (๐‘†) = (๐œ•๐‘† ๐‘“ )(0),
                                                 b                                                   (8)
                                   ๐œ•
where ๐œ•๐‘† =        ๐‘–โˆˆ๐‘† ๐œ•๐‘– and ๐œ•๐‘– = ๐œ•๐‘ฅ ๐‘– is the partial derivative.
              รŽ

   1Although the domain of ๐‘“๐œŒ is {โˆ’1, 1} ๐‘ , it only depends on the coordinates in free(๐œŒ).


                         T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                        6
          A S TOCHASTIC C ALCULUS A PPROACH TO THE O RACLE S EPARATION OF BQP AND PH

    Further, because ๐‘“ is multilinear, for any โ„Ž โˆˆ โ„ \ {0} and any standard basis vector ๐‘’ ๐‘– we
have
                                              ๐‘“ (๐‘ฅ + โ„Ž๐‘’ ๐‘– ) โˆ’ ๐‘“ (๐‘ฅ)
                                 (๐œ•๐‘– ๐‘“ )(๐‘ฅ) =                       .                        (9)
                                                      โ„Ž
    The following lemma is similar to [5, Claim A.5], which first appeared in [3] and [4, Claim
3.3].

Lemma 4.3. Let ๐‘“ : โ„ ๐‘ โ†’ โ„ be a multilinear polynomial. For any ๐‘ฅ โˆˆ [โˆ’1/2, 1/2]๐‘ , there exists a
distribution โ„› ๐‘ฅ over restrictions ๐œŒ โˆˆ {โˆ’1, 1, โˆ—} ๐‘ , such that for any ๐‘–, ๐‘— โˆˆ [๐‘],

                                            (๐œ•๐‘–๐‘— ๐‘“ )(๐‘ฅ) = 4 E            (๐œ•๐‘–๐‘— ๐‘“๐œŒ )(0) .
                                                                                   
                                                                                                                           (10)
                                                             ๐œŒโˆผโ„› ๐‘ฅ


Proof. We define โ„› ๐‘ฅ as follows: for each coordinate ๐‘– โˆˆ [๐‘] we independently set ๐œŒ ๐‘– to be 1 with
probability 14 + ๐‘ฅ2๐‘– , to be โˆ’1 with probability 14 โˆ’ ๐‘ฅ2๐‘– , and to be โˆ— with probability 12 .
   Using that ๐‘“ is a multilinear polynomial, and that        the coordinates are independent, we
deduce that for any ๐‘ฆ โˆˆ โ„ ๐‘ , ๐‘“ (๐‘ฅ + ๐‘ฆ) = E๐œŒโˆผโ„› ๐‘ฅ ๐‘“๐œŒ (2๐‘ฆ) . Then, using Equation (9),

          (๐œ•๐‘–๐‘— ๐‘“ )(๐‘ฅ) = ๐‘“ (๐‘ฅ + ๐‘’ ๐‘– + ๐‘’ ๐‘— ) โˆ’ ๐‘“ (๐‘ฅ + ๐‘’ ๐‘– ) โˆ’ ๐‘“ (๐‘ฅ + ๐‘’ ๐‘— ) + ๐‘“ (๐‘ฅ)
                                    ๐‘“๐œŒ (2๐‘’ ๐‘– + 2๐‘’ ๐‘— ) โˆ’ ๐‘“๐œŒ (2๐‘’ ๐‘— ) โˆ’ ๐‘“๐œŒ (2๐‘’ ๐‘– ) + ๐‘“๐œŒ (0) = 4 E            (๐œ•๐‘–๐‘— ๐‘“๐œŒ )(0) .
                                                                                                                  
                     = E                                                                                                     
                        ๐œŒโˆผโ„› ๐‘ฅ                                                                 ๐œŒโˆผโ„› ๐‘ฅ

    We now show our main result, which is analogous to [5, Theorem A.7] and [8, Theorem 2.4].

Theorem 4.4. Let ๐‘“ : {โˆ’1, 1} ๐‘ โ†’ {โˆ’1, 1} be a Boolean function, and let ๐ฟ > 0 such that for any
restriction ๐œŒ,                        ร•
                                          |b๐‘“๐œŒ (๐‘†)| โ‰ค ๐ฟ.
                                                    ๐‘†โІ[๐‘]
                                                     |๐‘†|=2

Let ๐›พ > 0 and let B be an ๐‘-dimensional Brownian motion with mean 0 and covariance matrix ๐พ.
Further assume that |๐พ ๐‘–๐‘— | โ‰ค ๐›พ for ๐‘– โ‰  ๐‘—.
    Let ๐œ€ > 0 and define the stopping time

                           ๐œ B min {๐œ€, first time that B๐‘ก exits [โˆ’1/2, 1/2]๐‘ }.

Then, identifying ๐‘“ with its multilinear representation, we have

                                           |E[ ๐‘“ (B๐œ )] โˆ’ E[ ๐‘“ (๐’ฐ๐‘› )]| โ‰ค 2๐œ€๐›พ๐ฟ.

Proof. First, we note that E[ ๐‘“ (๐’ฐ๐‘ )] = ๐‘“ (0). Note that B๐œ is always within [โˆ’1/2, 1/2]๐‘ . We can
apply Theorem 3.6

                                                   ๏ฃฎโˆซ ๐œ                                 ๏ฃน
                                                        1 ร•
                           E[ ๐‘“ (B๐œ )] โˆ’ ๐‘“ (0) = E ๏ฃฏ             ๐พ ๐‘–๐‘— (๐œ•๐‘–๐‘— ๐‘“ )(B๐‘  ) ๐‘‘๐‘  ๏ฃบ๏ฃบ .
                                                   ๏ฃฏ                                    ๏ฃบ
                                                   ๏ฃฏ                                                                       (11)
                                                   ๏ฃฏ 0 2 ๐‘–,๐‘—โˆˆ[๐‘]                        ๏ฃบ
                                                   ๏ฃฐ                                    ๏ฃป

                        T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                                                7
                                                      X INYU W U

Then, we use the upper bound ๐œ โ‰ค ๐œ€, and that (๐œ•๐‘–๐‘– ๐‘“ ) = 0 for all ๐‘– โˆˆ [๐‘] because ๐‘“ is multilinear,
to get

                                     ๏ฃฎ                                      ๏ฃน
                                              1 ร•
        | E[ ๐‘“ (B๐œ )] โˆ’ ๐‘“ (0)| โ‰ค ๐œ€ E ๏ฃฏ sup              ๐พ ๐‘–๐‘— (๐œ•๐‘–๐‘— ๐‘“ )(B๐‘  ) ๏ฃบ๏ฃบ
                                     ๏ฃฏ                                      ๏ฃบ
                                     ๏ฃฏ
                                     ๏ฃฏ๐‘ โˆˆ[0,๐œ] 2 ๐‘–,๐‘—โˆˆ[๐‘]                     ๏ฃบ
                                 ๐œ€๐›พ
                                     ๏ฃฐ              ร•                       ๏ฃป
                               โ‰ค         sup                 (๐œ•๐‘–๐‘— ๐‘“ )(๐‘ฅ)
                                   2 ๐‘ฅโˆˆ[โˆ’1/2,1/2]๐‘
                                                       ๐‘–โ‰ ๐‘—
                                                       ร•
                                                                       (๐œ•๐‘–๐‘— ๐‘“๐œŒ )(0)
                                                                                     
                               = 2๐œ€๐›พ        sup                E                          (Lemma 4.3)
                                        ๐‘ฅโˆˆ[โˆ’1/2,1/2]๐‘ ๐‘–โ‰ ๐‘— ๐œŒโˆผโ„› ๐‘ฅ

                                                         ๏ฃฎร•                   ๏ฃน
                                                                (๐œ•๐‘–๐‘— ๐‘“๐œŒ )(0) ๏ฃบ๏ฃบ
                                                         ๏ฃฏ                    ๏ฃบ
                               โ‰ค 2๐œ€๐›พ     sup        E ๏ฃฏ๏ฃฏ
                                     ๐‘ฅโˆˆ[โˆ’1/2,1/2]๐‘ ๐œŒโˆผโ„› ๐‘ฅ ๏ฃฏ ๐‘–โ‰ ๐‘—                ๏ฃบ
                                                         ๏ฃฐ                    ๏ฃป
                                                         ๏ฃฎ                    ๏ฃน
                                                         ๏ฃฏ ร•                  ๏ฃบ
                                                                      ๐‘“๐œŒ (๐‘†) ๏ฃบ๏ฃบ
                                                         ๏ฃฏ                    ๏ฃบ
                               โ‰ค 2๐œ€๐›พ     sup        E ๏ฃฏ๏ฃฏ              b                   (Equation (8))
                                     ๐‘ฅโˆˆ[โˆ’1/2,1/2]๐‘ ๐œŒโˆผโ„› ๐‘ฅ ๏ฃฏ๐‘†โІfree(๐œŒ)           ๏ฃบ
                                                         ๏ฃฏ                    ๏ฃบ
                                                         ๏ฃฐ   |๐‘†|=2            ๏ฃป
                               โ‰ค 2๐œ€๐›พ๐ฟ.                                                                     


5   Bound for classical circuits
We now construct the distribution ๐’Ÿ as described in Section 2.1. Due to Proposition 2.1, we can
take ๐’Ÿ to be the distribution defined by B๐œ , where B is a 2๐‘-dimensional Brownian motion
with covariance matrix ฮฆ, ๐œ is defined as in Theorem 4.4, and and ๐œ€ will be chosen appropriately
before Equation (12) below.
   Following Razโ€“Tal, we first use the following result of Tal [10] on the Fourier weight of
Boolean circuits.

Theorem 5.1 (Theorem 37(3) of [10]). There exists a constant ๐ถ > 0 such that the following holds.
Let ๐‘“ : {โˆ’1, 1}2๐‘ โ†’ {โˆ’1, 1} be a Boolean function computed by a Boolean circuit of depth ๐‘‘ and size
๐‘š > 1. Then for all ๐‘˜,            ร•
                                                  ๐‘“ (๐‘†)| โ‰ค 2(๐ถ(log ๐‘š)๐‘‘โˆ’1 ) ๐‘˜ .
                                                 |b
                                       ๐‘†:|๐‘†|=๐‘˜

    In particular, if ๐‘“ is computed by a bounded-depth circuit of quasipoly(๐‘) size, then
                                            ร•
                                                     ๐‘“ (๐‘†)| โ‰ค polylog(๐‘).
                                                    |b
                                          ๐‘†:|๐‘†|=2


    Since restriction does not increase circuit size or depth, we can apply Theorem 4.4 with

                        T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                              8
          A S TOCHASTIC C ALCULUS A PPROACH TO THE O RACLE S EPARATION OF BQP AND PH

๐œ€ = 1/(8 ln 2๐‘) and ๐›พ = โˆš1 , to deduce that
                               ๐‘

                                                                    polylog (๐‘)
                                         | E[ ๐‘“ (B๐œ )] โˆ’ ๐‘“ (0)| โ‰ค       โˆš       ,                 (12)
                                                                          ๐‘
where B๐œ is defined as in Theorem 4.4, justifying Equation (1).                                     


6    Quantum algorithm
Finally, we show that a 1-query ๐‘‚(log ๐‘)-time quantum algorithm can distinguish ๐’Ÿ from the
uniform distribution. This is virtually identical to the argument in [8, Section 6], but we can
again use some stochastic calculus tools on the stopping time built into our, slightly different,
distribution.
    The Forrelation query algorithm [1, 2] is an ๐‘‚(log ๐‘)-time quantum algorithm with inputs
๐‘ฅ, ๐‘ฆ โˆˆ {โˆ’1, 1} ๐‘ which accepts with probability (1 + ๐œ™(๐‘ฅ, ๐‘ฆ))/2, where
                                                            1
                                               ๐œ™(๐‘ฅ, ๐‘ฆ) B      h๐‘ฅ, ๐ป๐‘ ๐‘ฆi.                          (13)
                                                            ๐‘
When the pair (๐‘ฅ, ๐‘ฆ) is drawn from the uniform distribution, E[๐œ™(๐‘ฅ, ๐‘ฆ)] = 0. The quantum
algorithm is to prepare the uniform superposition over the basis states |1i . . . |๐‘i, query ๐‘ฅ, apply
the Walshโ€“Hadamard transform, query ๐‘ฆ, apply the Walshโ€“Hadamard transform again, then
measure in the computational basis and accept if the outcome is |1i. The quantum algorithm is
described in more detail in [1, Section 3.2].
    We show the following inequality [8, Claim 6.3], which shows that the quantum algorithm
distinguishes ๐’Ÿ from the uniform distribution with sufficiently high probability, justifying
Equation (2).
Proposition 6.1. E(x๐œ ,y๐œ )โˆผ๐’Ÿ [๐œ™(x๐œ , y๐œ )] โ‰ฅ 4๐œ€ .
Proof. We have
                                                       1
                              E         [๐œ™(x๐œ , y๐œ )] =  E[hx๐œ , ๐ป๐‘ y๐œ i]
                          (x๐œ ,y๐œ )โˆผ๐’Ÿ                  ๐‘
                                                       1
                                                     =   E[hx๐œ , ๐ป๐‘
                                                                  2
                                                                    x๐œ i] = E[kx๐œ k 2 ] .
                                                       ๐‘
By Proposition 3.5, we have
                                                   E[kx๐œ k 2 ] = E[๐œ] .                           (14)
By Markovโ€™s inequality,
                                               ๐œ€
                                                 Pr[๐œ > 2๐œ€ ].
                                                 E[๐œ] โ‰ฅ
                                               2
If ๐œ โ‰ค 2๐œ€ , it must be the case that the path exits [โˆ’1/2, 1/2]2๐‘ no later than 2๐œ€ . Hence, by the
union bound, we have
           Pr ๐œ โ‰ค 2๐œ€ โ‰ค 2๐‘ ยท Pr 1st coordinate of (x๐œ , y๐œ ) exits โˆ’ 12 , 12 not later than 2๐œ€ .
                                                                                        
                                                                                                  (15)

                         T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                      9
                                                          X INYU W U

                                                                                                    (1)
   Each coordinate of (x๐œ , y๐œ ) is a standard 1D Brownian motion since ๐พ ๐‘–๐‘– = 1 for all ๐‘–. Let B๐‘ก
denote the first coordinate of x๐‘ก . Applying Proposition 3.7,
                        "                         #
                                        (1)   1                                   1
                   Pr        sup      |B๐‘ก | โ‰ฅ         โ‰ค 2eโˆ’1/4๐œ€ = 2eโˆ’2 ln 2๐‘ โ‰ค        for ๐‘ โ‰ฅ 4.   (16)
                            0โ‰ค๐‘กโ‰ค๐œ€/2           2                                  4๐‘

Therefore, Pr[๐œ โ‰ค 2๐œ€ ] โ‰ค 12 , so E[๐œ] โ‰ฅ 4๐œ€ .                                                         


7    Acknowledgments
I would like to thank Ryan Oโ€™Donnell and Avishay Tal for helpful discussions and their
suggestions concerning an early draft. Thanks also to Gregory Rosenthal and anonymous
reviewers for helpful comments.


References
 [1] Scott Aaronson: BQP and the Polynomial Hierarchy. In Proc. 42nd STOC, pp. 141โ€“150.
     ACM Press, 2010. [doi:10.1145/1806689.1806711] 1, 2, 9

 [2] Scott Aaronson and Andris Ambainis: Forrelation: a problem that optimally separates
     quantum from classical computing. SIAM J. Comput., 47(3):982โ€“1038, 2018. Preliminary
     version in STOCโ€™15. [doi:10.1137/15M1050902] 2, 9

 [3] Boaz Barak and Jarosล‚aw Bล‚asiok: On the Raz-Tal oracle separation of BQP and PH, 2018.
     windowsontheory.org. 7

 [4] Eshan Chattopadhyay, Pooya Hatami, Kaave Hosseini, and Shachar Lovett: Pseudoran-
     dom generators from polarizing random walks. Theory of Computing, 15(10):1โ€“26, 2019.
     Preliminary version in CCCโ€™18. [doi:10.4086/toc.2019.v015a010] 7

 [5] Eshan Chattopadhyay, Pooya Hatami, Shachar Lovett, and Avishay Tal: Pseudorandom
     generators from the second Fourier level and applications to AC0 with parity gates. In
     Proc. 10th Innovations in Theoret. Comp. Sci. Conf. (ITCSโ€™19), pp. 22:1โ€“22:15. Schloss Dagstuhlโ€“
     Leibniz-Zentrum fuer Informatik, 2019. [doi:10.4230/LIPIcs.ITCS.2019.22, ECCC:TR18-155]
     7

 [6] Merrick Lee Furst, James B. Saxe, and Michael Sipser: Parity, circuits, and the polynomial-
     time hierarchy. Math. Systems Theory, 17(1):13โ€“27, 1984. Preliminary version in FOCSโ€™81.
     [doi:10.1007/BF01744431] 1

 [7] Bernt ร˜ksendal: Stochastic Differential Equations. Universitext. Springer, 6th edition, 2003.
     [doi:10.1007/978-3-642-14394-6] 3, 4, 5

                        T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                       10
         A S TOCHASTIC C ALCULUS A PPROACH TO THE O RACLE S EPARATION OF BQP AND PH

 [8] Ran Raz and Avishay Tal: Oracle separation of BQP and PH. In Proc. 51st STOC, pp. 13โ€“23.
     ACM Press, 2019. [doi:10.1145/3313276.3316315, ECCC:TR18-107] 1, 2, 3, 5, 6, 7, 9

 [9] Daniel Revuz and Marc Yor: Continuous Martingales and Brownian Motion. Volume 293 of
     Grundlehren der Math. Wiss. Springer, 3rd edition, 1999. [doi:10.1007/978-3-662-06400-9] 5

[10] Avishay Tal: Tight bounds on the Fourier spectrum of AC0 . In Proc. 32nd Comput. Complexity
     Conf. (CCCโ€™17), pp. 15:1โ€“15:31. Schloss Dagstuhlโ€“Leibniz-Zentrum fuer Informatik, 2017.
     [doi:10.4230/LIPIcs.CCC.2017.15, ECCC:TR14-174] 5, 8


AUTHOR

     Xinyu Wu
     Graduate student
     Computer Science Department
     Carnegie Mellon University
     Pittsburgh, PA, USA
     xinyuwu cmu edu
     https://www.andrew.cmu.edu/user/xinyuw1/


ABOUT THE AUTHOR

     Xinyu Wu is a Ph. D. student at Carnegie Mellon University, advised by Ryan
        Oโ€™Donnell and Pravesh Kothari. Her research interests are in spectral graph
        theory, free probability, and their applications in understanding average-case
        problems and quantum computing. She grew up in Singapore, and did her
        undergraduate degree in math and computer science also at Carnegie Mellon.
        She also enjoys cooking, cycling, and cat videos.




                    T HEORY OF C OMPUTING, Volume 18 (17), 2022, pp. 1โ€“11                    11